Why are half-band decimation filters and studio sample-rate conversion a problem, and how much aliasing rejection is enough?

Short version

Watts says the aliasing in modern recordings enters at the ADC's decimation filter. He describes that filter as a half-band type, only about 6 dB down at half the sample rate, so material above that point folds back into the audio band as aliasing Head-Fi, Feb 2016Head-Fi, Jan 2018. He says this is audible because it degrades the timing of transients, even though the aliasing sits at frequencies one cannot hear on their own Head-Fi, Apr 2016Head-Fi, Jun 2019Distorsiuni Interioare 56:10. He says non-integer sample rate conversion has a similar effect, adding distortion, noise and timing errors Head-Fi, Apr 2019Head-Fi, Jun 2019Head-Fi, Sep 2020.

On how much aliasing rejection is enough he has given a series of dated results rather than a threshold: 140 dB described as a "sledge hammer approach" for Hugo's DSD filter (2014), 140 dB not enough (2015 to 2016), the Hugo 1 DSD filter audible and -220 dB without loss in his later DSD filters (2018 to 2019), -350 dB audible in an ADC sample rate converter (2021), and in July 2024 the limits still being determined Head-Fi, Apr 2014Head-Fi, Apr 2016Head-Fi, Jun 2018Head-Fi, Aug 2018Head-Fi, May 2021Head-Fi, Jul 2024.

What he says is wrong with half-band decimation filters

In February 2016 he wrote that "normal ADC decimation filters are half band, so offer worst case only -6dB rejection" Head-Fi, Feb 2016. In April 2016 he described the conventional defense ("you can't hear anything from 20 to 20k, so it does not matter") and answered that "the aliasing has completely destroyed the timing of transients" Head-Fi, Apr 2016.

In August 2017 he said ADC chip designers "allow unacceptable levels of aliasing as they all use cheap half-band filters" Head-Fi, Aug 2017. In February 2018 he said that existing decimation filters, for conversions such as 96k to 48k, use cheap half-band filters that create aliasing in the 20 to 24 kHz band, and that "this damages timing too" Head-Fi, Feb 2018. In June 2019 he said the aliasing components near 20 to 24 kHz are not audible by themselves, but "they will change the timing of transients, and this is audible" Head-Fi, Jun 2019.

In an October 2026 podcast he said half-band filters are used because they are cheap and give a smaller silicon die, that the response is only -6 dB at half the sample rate (22.05 or 24 kHz), and that the reflections continue "right down to DC, although not at minus 6 dB, we'll be at minus one hundred and twenty or something like that", with most of the aliasing errors between about 18 and 24 kHz Distorsiuni Interioare 54:40Distorsiuni Interioare 55:24. A podcast dated three days later has him putting aliasing products at around minus 120 dB for the best A to D converters within the audio bandwidth Part-Time Audiophile 57:21.

He also describes an earlier stage in a conventional ADC. In October 2017 he said a standard ADC takes the noise shaper output at about 3 MHz and decimates it to 384 or 768 kHz with a moving-average (sinc) filter, with attenuation of "around 20 dB" miceblue425 36:03miceblue425 36:50. In February 2019 he wrote that the decimation in conventional filters is poor, with "20 dB or so rejection" Head-Fi, Feb 2019.

On whether the damage can be undone, he made the same point in posts from 2017 to 2022. In October 2017 he wrote that "no filter can remove aliasing" Head-Fi, Oct 2017. In January 2018: "with aliasing the damage has already been done, and no interpolation filter can repair that damage" Head-Fi, Jan 2018. In May 2022: "once the damage is done, it can't be recovered by subsequent filtering" Head-Fi, May 2022. In July 2022, after testing a filter on the 20 to 22.05 kHz band after reconstruction (at 705 kHz after WTA1), he found little or no benefit and wrote "I guess once the aliasing damage is done within the ADC decimation filter, then it is impossible to repair that damage" Head-Fi, Jul 2022. In July 2018 he said careful listening tests showed that "the sinc function type response sounded much better, even when using modern recordings that have large amounts of ADC aliasing because they use half band filters", compared with what one might expect from a filter that removed everything above 20k Head-Fi, Jul 2018.

He has linked this to older recordings, with hedges. In October 2018 he wrote "I suspect that the reason 1960's recordings often sound amazing on the M scaler is down to the proper anti aliasing that was incorporated when the recording was eventually digitised" Head-Fi, Oct 2018. In August 2023 he contrasted the Sony PCM1630's passive anti-aliasing filter ("better than 50dB rejection at 22kHz") with today's half-band decimation at -6 dB Head-Fi, Aug 2023, and on early Decca recordings wrote "we can't be certain about this", saying an EQ or high-frequency distortion issue could also explain the effect and that he suspected both factors were at play, with aliasing the dominant part Head-Fi, Aug 2023.

Sample rate conversion

In June 2014 (Hugo thread) he reported a track that sounded awful because the PC was converting everything to 48 kHz: "all the Hugo magic had gone" Head-Fi, Jun 2014. In September 2014 he said that if the USB input could not be used, converting 88.2 to 44.1 would probably sound best, because good quality decimation is easier than raising the rate and "there are no timing problems with decimation, its just a straight forward anti-aliasing filter" Head-Fi, Sep 2014. In May 2017 he said down-sampling "creates transient timing errors through aliasing; but the up-sampling process creates many times more errors" Head-Fi, May 2017. The 2014 and 2017 wordings about decimation differ, and he did not comment on the difference in these items.

He made the same point about non-integer conversion repeatedly from 2015 to 2024. In March 2015 he said that if a 96 kHz master was taken to 44.1 kHz with a poor sample rate converter, "no amount of increased taps will solve the SRC distortion and noise" Head-Fi, Mar 2015. In January 2016 he said conventional sample rate converters "generate large amounts of THD and noise, in particular slew related noise floor modulation" Head-Fi, Jan 2016. In January 2017 he said decimating 192k to 48k "should be relatively transparent", while "Non integer SRC is very tough to do transparently" Head-Fi, Jan 2017. In February 2018 he wrote "SRCs are impossible to make without measurable THD and noise issues, and on their own create timing issues too" Head-Fi, Feb 2018. In June 2018 he said he was bothered by sample rate conversion, "which adds lots of distortion and noise and degrades transient timing", and not much by a lower rate "so long as it is binary multiples" Head-Fi, Jun 2018.

In April 2019: "Non integer sample rate conversion (SRC) adds a lot of THD and noise and too degrades timing" Head-Fi, Apr 2019. In June 2019 he said it is "impossible to do it without adding distortion" and that the added noise is slew-rate related, giving noise floor modulation that depends on high-frequency content Head-Fi, Jun 2019. In September 2020 he said he buys a 96/24 file over a 44.1/16 one because the sample rate conversion process "creates timing errors and distortion and noise" Head-Fi, Sep 2020.

In March 2023 he said that "No amount of WTA filtering will recover the damage done from non integer sample rate conversion" Head-Fi, Mar 2023. In August 2023 he compared Arrival by Johann Johannsson at 16-bit 44.1k with a 48k 24-bit version, said the difference was extraordinary, and wrote "It just shows how much damage non integer sample rate conversion from 48k to 44.1k does" Head-Fi, Aug 2023. In January 2024, answering a video's claim that SRC is so good you cannot hear it, he called that wrong, wrote "To make SRC work completely transparently is almost impossible", and advised giving his DACs bit-perfect data with no SRC or upsampling Head-Fi, Jan 2024.

For integer ratios he was milder. In August 2018 (TT2 thread) he said "keep to integer decimation (192>96) and the loss will be minimal with a decent filter" Head-Fi, Aug 2018.

Inside the ADC itself, in January 2021 he said studios run a common clock, so going from his local 104 MHz clock to 768 kHz synced to an external clock needs a very complex fractional interpolation filter, and that "the work suggests that aliasing, if it's observable on a digital domain measurement, is easily audible" even in single-blind tests; he said he would not publish results yet Head-Fi, Jan 2021. In July 2021 he described the same converter's time resolution creating noise floor modulation, still audible from atto to zepto second resolution, and concluded that "the target for noise floor modulation needs to be around -400dB"; that figure is for noise floor modulation, not aliasing Head-Fi, Jul 2021. In September 2023 he said the stereo ADC then being designed would test an SRC clocking option for syncing to external word clocks "with (hopefully) no loss in sound quality" Head-Fi, Sep 2023.

How much aliasing rejection is enough

2014: Hugo's 140 dB DSD filter

In April 2014 he explained that Hugo's DSD filter had to decimate to 16FS, and that he "decided on a sledge hammer approach to aliasing problems, by having 140 dB of rejection", which he said was much better than pro standard ADC aliasing filters. Compared with the Qute filter, he said the Hugo filter sounded very much better, "with no detectable loss in transparency or timing" Head-Fi, Apr 2014. His later statements about the same Hugo 1 filter are in the 2018 to 2020 section below.

2015 to 2016: 140 dB not enough

In February 2015 he said that an ADC anti-aliasing filter at -144 dB (24-bit accuracy) would need about 2K taps or more, that ADC chips have significant aliasing, and that listening would "find out what stop band attenuation is really needed" Head-Fi, Feb 2015. In March 2015 he said decimation done correctly with a very long tap WTA filter "should be completely transparent", to be established after listening tests with the pro audio ADC then under development Head-Fi, Mar 2015. In a May 2015 interview about DAVE he said he had been working with decimation accuracies of 140 dB on a filter and that "it actually turned out that's not good enough", because non-decimating versions sounded noticeably better, and that his pro audio ADC would need better than 140 dB; the passage does not name the filter HIFICLUBE.NET JVH 3:52HIFICLUBE.NET JVH 4:50.

In February 2016 he wrote that he had used -140 dB decimation filters "and can still hear the effects of aliasing" Head-Fi, Feb 2016. In April 2016 he wrote "I know that -140 is not enough" and asked whether -160 or -200 dB would be Head-Fi, Apr 2016. In May 2016 he wrote that 140 dB rejection "is not good enough in that it changes the SQ and gives measured artifacts", and asked "how much is good enough?" Head-Fi, May 2016.

2017 to 2019: the figures he quotes for Davina

For his own ADC project (Davina) he quoted different design figures at different times: two filters at 300 dB at and above half the sample rate (August 2017), a first filter with 260 dB rejection (October 2017), 240 dB "from my memory of my decimation filter design" (January 2018), aliasing "below -240 dB or so" (February 2018), more than 250 dB with aliasing guaranteed below -250 dB in a simulation (September 2018), 250 dB (March 2019) and 220 dB (June 2019) Head-Fi, Aug 2017Head-Fi, Oct 2017Head-Fi, Jan 2018Head-Fi, Feb 2018Head-Fi, Sep 2018Head-Fi, Sep 2018Head-Fi, Mar 2019Head-Fi, Jun 2019. In July 2018 he said the 705/768 to 44.1/48 kHz filters had been tested with "absolutely no aliasing (to 24 bit rounding error)" Head-Fi, Jul 2018, and in August 2018 that the 768 to 48 kHz filters were tested and that from 384 kHz to 104 MHz rejection was better than 250 dB Head-Fi, Aug 2018. In October 2018 he wrote that he had to re-design the filter from scratch "as it failed on testing" Head-Fi, Oct 2018. These are design values he reported, and the differences between them are not a stated change of view.

In July 2018 he wrote that it is "perfectly possible to bandwidth limit and decimate without any aliasing (unless -240 dB is important!)" Head-Fi, Jul 2018. In March 2019 he said stop band attenuation matters little beyond about -150 dB for interpolation filters, but for decimation filters "it is much more important" Head-Fi, Mar 2019.

2018 to 2020: the DSD filter experiments (Hugo, Hugo 2, TT2, M scaler, DAVE)

He said he used DSD-to-PCM decimation filters to find out how much aliasing matters, because it bore on the ADC project Head-Fi, Jan 2018Head-Fi, Aug 2018Head-Fi, May 2019. In January 2018 he said the Hugo 1 filter had aliasing "you could see stuff at -150dB, and this was surprisingly audible", while the newer filter had aliasing better than -240 dB Head-Fi, Jan 2018. In June 2018 he described the Hugo 1 filter as having aliasing better than -130 dB, said this "was found to be audible", which led to DAVE's non-decimating DSD+ mode, and concluded "-130 dB is simply not good enough at all" Head-Fi, Jun 2018. The filter in Hugo 2, TT2 and the M scaler has -220 dB aliasing, which he said sounds better than DSD+ even though it decimates, attributing that to more effective RF noise reduction; he wrote "my target is to always exceed 200dB performance, and this I do on all decimation filters I do" Head-Fi, Jun 2018.

In August 2018 he said decimating to 16FS had no sound quality loss "so long as the aliasing is kept to levels below -220 dB" Head-Fi, Aug 2018. In February 2019 he wrote that the M scaler filter showed "one can successfully decimate to 16FS without any losses in SQ so long as there is zero aliasing", with even small amounts of aliasing subjectively important Head-Fi, Feb 2019. In May 2019 he gave Hugo 1 as -130 dB rejection with worst-case aliasing at -150 dB and Hugo 2 as -220 dB with worst case -240 dB, wrote "very tiny amounts of aliasing is audible", and said decimation to 705.6 or 768 kHz is not a problem "so long as you make sure any aliasing at all is at -240dB or better" Head-Fi, May 2019. In October 2020 he again tied the absence of a decimation penalty to 220 dB filtering Head-Fi, Oct 2020. These statements concern DSD decimation in his DACs; he presents them as evidence for the ADC question, not as measurements of studio ADCs.

2021: -350 dB aliasing audible

In May 2021 he described testing the sample rate converter for the ADC project. A first fractional Lagrange interpolator had worst-case aliasing at -200 dB, and he said turning it off (then getting -160 dB aliasing) was easily audible; he kept improving it over several months to below -350 dB Head-Fi, May 2021. He wrote that overlaying waveforms with -310 dB aliasing against -350 dB, "it would be impossible to see the difference", yet "we can actually perceive these very, very small differences", and he described a blind listening test with his son whose assessment agreed with his own Head-Fi, May 2021. Four days later he wrote that his decimation filters were -240 dB and that "work on the ADC sample rate converter indicates that aliasing at -350dB is audible - so I now have to re-do those filters" Head-Fi, May 2021.

2022 to 2024: no threshold, still being determined

In May 2022 he wrote that "defining acceptable aliasing performance is not easy, as the brain perceptually is extremely sensitive to vanishingly small amounts of aliasing", and that so long as acceptable aliasing performance is achieved it does not matter in principle whether the decimation filter is IIR or FIR Head-Fi, May 2022. The next day he added that an IIR filter is fine for decimation "so long as it has infinite attenuation at 24kHz and above" Head-Fi, May 2022. In March 2024 he said Quartet creates its 176.4/192 kHz output by decimation from 705.6/768 kHz with decimation filters "like nothing ever achieved before in both accuracy, capability and transparency" Head-Fi, Mar 2024. In May 2024 he said he had "created error free filters for anti aliasing" and that ADC decimation requirements are very different from reconstruction Head-Fi, May 2024. In July 2024, after discussing the M scaler DSD filter (better than 200 dB rejection), he said decimation to 16FS is fine with adequate attenuation and that "we have to do better than 200dB attenuation", but that this "doesn't tell us the limits" for subjectively perfect decimation, which was "something I am determining at the moment with the ADC project" Head-Fi, Jul 2024.

2026: Quartet and the pending comparison

In a Soundnews interview recorded in autumn 2026 (published October 2026) he said the Davina design had been scrapped because components were discontinued, and that he then had to finish the ADC when Chord wanted an analog input in Quartet article, Oct 2026article, Oct 2026article, Oct 2026; in a podcast published in October 2026 he said Quartet has a pulse array ADC that he had to perfect to finish the project Distorsiuni Interioare 15:29. He said Quartet has an 18 kHz filter used as part of the decimation process within the ADC that eliminates everything above 20 kHz article, Oct 2026, and that Quartet lets him "completely eliminate these aliasing ish issues" that standard pro ADCs have Distorsiuni Interioare 13:35. In the podcast three days later, speaking of his pulse array ADC (the transcript reads "my Pro A D C"), he said the aliasing issue is eliminated, with the aliasing numbers "vanishingly small" Part-Time Audiophile 57:48. He also reported that a female vocalist "sounded out of tune on the commercial ADC and in tune on Quartet" article, Oct 2026. In these items he gives no dB figure for Quartet's ADC decimation, and he does not say that aliasing was isolated as the only difference between the two ADCs.

He said 768 kHz recordings can now be decimated to 48 kHz and brought back to 768 kHz to see what is lost, with the objective of a difference "so small that we don't need to worry about it" article, Oct 2026. Asked what remains open, he said the key question is how much difference that round trip makes and whether there is "more work needed on decimation and/or interpolation", to be answered as a bank of test recordings is built "now and into 2027" article, Oct 2026.

What he has not said

Watts has not given a single dB figure as the answer to how much aliasing rejection is enough. He has not published the results of the 768 to 48 to 768 kHz comparison and in a written answer published in October 2026 described it as still to come article, Oct 2026. He has not given a dB figure for Quartet's ADC decimation filters, and in July 2024 he said the limits for subjectively perfect decimation were still being determined Head-Fi, Jul 2024.

Quotes

“Aliasing is a major problem - normal ADC decimation filters are half band, so offer worst case only -6dB rejection.”2016-02-26 Head-Fi, Feb 2016

“Non integer sample rate conversion (SRC) adds a lot of THD and noise and too degrades timing.”2019-04-25 Head-Fi, Apr 2019

“And the answer was - very tiny amounts of aliasing is audible.”2019-05-25 Head-Fi, May 2019

“defining acceptable aliasing performance is not easy, as the brain perceptually is extremely sensitive to vanishingly small amounts of aliasing”2022-05-09 Head-Fi, May 2022

“It doesn't tell us the limits that are requred for subjectively perfect decimation. That's something I am determining at the moment with the ADC project.”2024-07-01 Head-Fi, Jul 2024

Dated positions

Caveats and open points

• No single dB figure is given by Watts as the answer to how much aliasing rejection is enough; the dated results (140 dB not enough, the Hugo 1 DSD filter audible, -220 dB without loss in DSD filters, -350 dB audible in an ADC sample rate converter) came from different tests and different filters and are not one threshold.
• His figures for the Hugo 1 DSD filter vary by item: 140 dB of rejection (April 2014), aliasing visible at -150 dB (January 2018), aliasing better than -130 dB (June 2018), and -130 dB rejection with worst-case aliasing at -150 dB (May 2019). In 2014 he reported no detectable loss against the Qute filter; from 2018 he described its aliasing as audible against his newer filters. He did not comment on the difference in these items.
• The 2018 to 2020 figures (-130, -220, -240 dB) are for DSD-to-PCM decimation in Hugo 1, Hugo 2, TT2, the M scaler and DAVE, not for studio ADC half-band filters; he presents them as evidence for the ADC question.
• The Davina design figures (300, 260, 240, 250, 220 dB) are values he reported at different times, one of them from memory; their differences are not a stated change of view.
• The May 2015 interview does not name the filter that had 140 dB accuracy. The 2015 and 2017 interview figures and the October 2026 figures (-6 dB, about -120 dB) come; readers should check the recordings.
• The 2021 listening results are his own tests, including a blind test with his son; the -350 dB result concerned the ADC sample rate converter, and the conclusion about the decimation filters is his inference, stated as what the work 'indicates' in one forum post. The July 2021 -400 dB figure is a target for noise floor modulation, not aliasing.
• His 2014 remark that decimation has no timing problems beyond a straightforward anti-aliasing filter and his later remarks that half-band decimation aliasing degrades timing are reported side by side; he did not comment on the difference in these items.
• His links between aliasing and the sound of 1960s or Decca-era recordings are hedged by him (suspected, not certain).
• The Soundnews piece is an edited published interview; the Quartet statements do not isolate aliasing as the only difference between Quartet's ADC and the commercial ADC he compared against. In the October 14, 2026 podcast the transcript reads 'my Pro A D C' where the context is his pulse array ADC.
• No dB figure for Quartet's ADC decimation filters was found.

Topics: Decimation, aliasing and sample-rate conversion, Davina and ADC design, Recordings, mastering and the production chain